[Paper Review] Global Flood Prediction: a Multimodal Machine Learning Approach

Cynthia Zeng, Dimitris Bertsimas|arXiv (Cornell University)|Jan 29, 2023

Flood Risk Assessment and Management6 citations

TL;DR

This paper proposes a multimodal machine learning framework that combines tabular historical flood data and text-based geographical descriptions from Wikipedia to predict global flood risk over 1–5 year horizons. Using transfer learning with DistilBERT to extract contextual embeddings from geographical text, the model achieves 75%–77% ROCAUC, outperforming single-modality and baseline models, demonstrating the value of integrating textual geospatial context for long-term disaster forecasting.

ABSTRACT

Flooding is one of the most destructive and costly natural disasters, and climate changes would further increase risks globally. This work presents a novel multimodal machine learning approach for multi-year global flood risk prediction, combining geographical information and historical natural disaster dataset. Our multimodal framework employs state-of-the-art processing techniques to extract embeddings from each data modality, including text-based geographical data and tabular-based time-series data. Experiments demonstrate that a multimodal approach, that is combining text and statistical data, outperforms a single-modality approach. Our most advanced architecture, employing embeddings extracted using transfer learning upon DistilBert model, achieves 75\%-77\% ROCAUC score in predicting the next 1-5 year flooding event in historically flooded locations. This work demonstrates the potentials of using machine learning for long-term planning in natural disaster management.

Motivation & Objective

• To develop a global, multi-year flood risk prediction system that extends beyond short-term physical models.
• To investigate whether integrating text-based geographical descriptions improves long-term flood prediction accuracy.
• To evaluate the effectiveness of multimodal learning—combining statistical and textual data—compared to single-modality approaches.
• To demonstrate the generalizability of the framework to other natural disasters such as wildfires and droughts.
• To provide a scalable, data-driven tool for long-term disaster preparedness and infrastructure planning.

Proposed method

• The framework processes two data modalities: tabular time-series data from the GDIS and EM-DAT datasets, and text-based geographical descriptions from Wikipedia.
• Geographical text is extracted using the Wikipedia-API, with synonym matching to resolve location name variations.
• Text embeddings are generated using DistilBERT, with three variants: (1) original frozen model, (2) fine-tuned on flood frequency, and (3) transfer learning with a 32-dimensional projection head.
• Statistical features include 24 variables such as flood count, damage cost, and year, with one-hot encoding for disaster types.
• Multimodal predictions combine statistical features and text embeddings via early fusion in a feedforward neural network.
• Model training uses binary classification with positive labels for locations with >2 historical floods, and evaluation is performed via out-of-sample ROCAUC, accuracy, F1, and balanced accuracy.

Experimental results

Research questions

• RQ1 Can integrating text-based geographical descriptions improve long-term global flood risk prediction beyond traditional statistical models?
• RQ2 How does a multimodal machine learning approach combining tabular and textual data compare to single-modality models in predicting 1–5 year flood risk?
• RQ3 Does fine-tuning or transfer learning on pre-trained language models like DistilBERT enhance flood prediction performance on geospatial text?
• RQ4 To what extent does the inclusion of textual context reduce prediction error compared to a deterministic baseline that assumes current risk persists?
• RQ5 Can this framework be generalized to other natural disaster types such as droughts or wildfires?

Key findings

• The multimodal model using transfer learning with DistilBERT achieved 75%–77% ROCAUC for 1–5 year flood risk prediction, significantly outperforming single-modality models.
• The best-performing multimodal architecture improved the baseline model by 42 percentage points in ROCAUC and 25 percentage points in balanced accuracy.
• Single-modality models using only statistical data outperformed the baseline by 35 percentage points in ROCAUC, indicating inherent predictive power in historical data.
• Fine-tuning or direct use of pre-trained DistilBERT did not improve performance over single-modality baselines, highlighting the necessity of transfer learning with task-specific adaptation.
• The framework generalizes well to other disaster types, suggesting broad applicability in long-term natural disaster forecasting.